As Long as No Rate-Hike Cycle Begins, Is the Current Pullback a Buying Opportunity?

Author: Mr. Z, 168X

Macro is just noise; sentiment is the signal: In June, mindlessly piling into hardware was euphoria; in late August, nobody wanted semiconductors—that's desolation.

Taipei time, September 9, 2026, Wednesday, 10:00 AM. Last Friday's nonfarm payrolls came in at 162,000, far above the expected 50,000; this Friday CPI is on deck. With no US-Iran reconciliation, oil prices are hovering between $90 and $100, and the market is pricing in 1.5 rate hikes this year. Semiconductors went through a July decline without any rebound, a small rebound in the first half of August, and a second test of lows in the second half. Today, hardware rebounded from lows while software fell—the market is still running the "software down, hardware up" narrative. Is rate hikes the main theme? Should you buy hardware or software?

This episode of 168X Strategy Room welcomes back our old friend, Investment Talk (@TJ_Research). In this episode, he puts macro back in its place as noise: as long as we don't enter a rate-hike cycle, the market can digest two or three rate hikes. CPI itself doesn't matter; the Fed's reaction function is what matters. The market's expectations for Nvidia have been wrong since 2023—depreciation mismatch is a false proposition, and chips from three years ago now pay for themselves in a year. Open-source models are handing pricing power back to Cloud Service Providers (CSPs); the era when Amazon only made $4 on a $100 API call is about to change. Token price declines are the prerequisite for an agent explosion. One message for investors: Buy when no one cares, sell when everyone's excited. This year's market has already tested us many times!

1. Rate Hikes Aren't the Main Theme: As Long as No Rate-Hike Cycle Begins, Two or Three Hikes Are Just Speed Bumps

Mr. Z: Welcome everyone to 168X. Today is Taipei time, Wednesday, September 9, 10:00 AM. It's an honor to have our old friend Investment Talk back. Let's start with the macro picture: last week's nonfarm payrolls, Friday night's CPI, whether rates will be hiked, and Waller's comments.

Investment Talk: Last Friday's nonfarm payrolls and unemployment rate, from any angle, were a set of better-than-expected numbers. At the very least, they show the labor market is healthy, with no signals pointing to deterioration. Of course, from the bond market's trading results, the market didn't significantly raise the probability of a September hike after the data, mainly because from the Fed's perspective, they have two mandates, and the labor market is usually not the focus relative to CPI: for now, it's stable—not getting much worse or much better—and they won't treat one month's data as a trend. So the main factor deciding whether to hike in September is still this Friday's CPI.

Today someone asked me whether to hedge or adjust positions ahead of Friday's CPI. Personally, I won't. Even if they hike this time—September and then December—we still won't enter a rate-hike cycle. As long as we don't enter a rate-hike cycle, whether it's two hikes or three, the ultimate impact on asset prices won't be very significant. We are not in a rate-hike cycle like 2022, at least not based on the current inflation trend. So rate hikes themselves are not a major theme; they're just noise, a speed bump on the path of fundamentals and AI. Another point: today the index closed lower, and overall market participation and sentiment, looking at the S&P, are not high. Entering Friday's CPI with low sentiment, there are only two scenarios: if the data is bad, a market decline is reasonable, but the downside is limited because sentiment is already low; if the data is good or in line with expectations, a rebound in a low-participation, low-sentiment environment is very reasonable. If it's bad, it falls a bit, and after next week's rate hike lands, the market may stabilize and rebound. So given the current sentiment and participation, there's no great need to hedge against CPI.

2. CPI Itself Doesn't Matter: In 2021, CPI Rose from 2% to 5%, and the Market Kept Rising Until November

Mr. Z: With no US-Iran reconciliation and oil prices hovering between $90 and $100, how big is the impact on CPI? It looks like Friday's number will likely be ugly.

Investment Talk: I've shared a view before: CPI is not important; what matters is how the Fed interprets CPI. What drives market movements is not the CPI data itself, but the Fed's reaction function. If CPI is bad but the Fed says there's no need to hike, there's no impact on asset prices. The best proof is 2021: CPI rose from 2% to 5%, but the Fed remained dovish, and asset prices kept rising until November 2021, when the Fed pivoted. So whether CPI is good or bad is not the direct cause of asset price moves. The market reacts immediately to the data, but the reaction function is more important. This time, whether they hike twice or three times, whether CPI is good or bad—if it's bad, they'll ultimately hike 2.5 times. From a 3-to-6-month perspective, will two hikes versus three really have a huge impact on asset prices? Not that much.

Building on that: oil prices are part of CPI. If CPI is not that important relative to the Fed's reaction function, then oil prices are an even less important factor for CPI. Moreover, the Fed focuses more on core CPI, which at least strips out oil prices. Recently I've heard a different view: if there are signals of de-escalation, end, or pause in the US-Iran war, oil prices fall, CPI falls, and the Fed might even cut rates this year. I don't quite agree. Oil prices have no direct impact on core CPI, and in the short term, from a transmission perspective, there won't be much impact either. So the end of the US-Iran war won't have a huge impact on core CPI, core PCE, or Fed policy. Of course, if it truly ends, it would help reduce the urgency to hike, but it's not a very important factor.

3. Two-Year Treasury Yields: Standing with the Market, 1.5 Hikes This Year

Mr. Z: You had a post I found interesting: data from past years shows that rising two-year Treasury yields are a leading indicator, with a high probability that the Fed's rate range will be raised. There are signs of that this time too.

Investment Talk: Yes. If I had to guess whether the Fed will hike this year, I'd agree with the market: whatever the market does, I agree that's what the Fed will ultimately do. The market is pricing in 1.5 hikes this year, so I think there will be 1.5 hikes this year. It's not that every time two-year yields rise, the Fed will definitely hike. The market can be wrong, but being wrong requires data to slap the market's expectations—for example, if inflation data is very good and employment data is terrible, the market will naturally push two-year yields down. The bottom line is: when the market is pricing in rising two-year yields and potential hikes, at least the Fed cannot cut rates. Ultimately, it falls into one of two options: hold steady or hike. So I stand with the market. If the market thinks there will be hikes, I'll make investment decisions as if there will be hikes. I won't bet against the market by positioning for no hikes this year.

Because at least right now, I don't see any signs of core inflation or core PCE slowing down. In the medium to long term—at least 6 months to a year or even longer—I don't see signs of CPI slowing: AI capex is being poured in, overall wage growth hasn't returned to pre-pandemic levels and is still higher than pre-pandemic, and much consumption remains resilient. Inflation will be very hard to get back to 2%. So the best case is no hikes, the worst case is two to three hikes—it's just moving within those two options. In my investment framework, over the next 3 to 6 months, macro itself is still mostly noise: when the market is doing very well, it needs some negative news to sound a warning bell. But stepping back, as long as we don't enter a rate-hike cycle, the market can digest two or three hikes. Two or three hikes are a result; whether it's driven by inflation or something else in between doesn't really matter.

4. Extremes in Sentiment: In June, Everyone Debated Whether Micron Was a Cyclical Stock; in Late August, Nobody Wanted Semiconductors

Mr. Z: I saw you recently said that in June, deciding to buy semiconductors and sell software sounded like a stupid move; now it's kind of the opposite—everyone should gradually add to hardware/semis and slowly sell software. Is my interpretation correct?

Investment Talk: Half right. I didn't say software needs to be reduced; that tweet was just a description of sentiment. At the end of June, the market was debating whether Micron was a cyclical stock—very heated debate. Should Micron, as a cyclical, get a 5x P/E, or more than 10x, 20x? If you had to pick a starting point for this whole market pullback, it might be after Micron's earnings report, when the entire hardware and semiconductor complex started falling. But at that time, nobody was bullish on software; it was seen as a sector being disrupted by AI—sentiment was extremely extreme. By late August, sentiment had turned very pessimistic on semiconductors: semis went through a July decline without any rebound, a small rebound in the first half of August, then another decline in the second half. The second test of lows is the most painful—some names approached their late-July lows, and many chip and semi holders were very discouraged. But that's exactly when market sentiment is very low: SOX's forward P/E is close to late July. If I thought late July was cheap, then late August, with valuations close to late July, is still cheap. In mid-to-late June, semiconductor valuations were 28x, 29x, 30x, and everyone was debating whether Micron was a cyclical and chasing semis. Now, AI is still advancing with no signs of slowing, valuations have fallen 40% from the peak, and nobody has any confidence in semis. From sentiment alone, you can feel it: you shouldn't follow sentiment. That was my main point.

5. Software's Right Side Has Arrived: The First Quarter of Falsification

Investment Talk: As for software, I didn't say it should be sold. In the last couple of days, semis rebounded from lows while software fell again—the market is still running the "software down, hardware up; hardware down, software up" story. I actually think software will have a good entry point on the right side; the left side should already be behind us. Because from this quarter's earnings, this is the first quarter where many software names showed in their reports that the integration of their business with AI is starting to monetize. AI's implementation in software is just beginning. So the "AI disrupts software" narrative that ran for 8 quarters—over two years—will not only be falsified, but the market may slowly wake up: software valuations shouldn't be 10x or 11x; they should at least return to around 20x, or 70-80% of previous valuations. I'm not saying you should sell software now, but at least you shouldn't sell hardware. On the contrary, hardware—semiconductors—are still cheap.

6. Buy SOX Without Thinking: Qualcomm Is a Good Left-Behind Name

Mr. Z: Within semis—foundry, GPU, optical communications, memory—which sector should people look at first?

Investment Talk: Any of them. Look at which names have rebounded and which have pulled back. Optical still has some strong ones, like Lumentum near new highs—it surged yesterday. But not all optical names are strong; there's divergence among individual stocks. The simplest approach, if you don't want to think, is to just add the index—SOX is cheap. If you have to differentiate, before coming in today I was looking at the Qualcomm-Amazon partnership. I think Qualcomm is a very good name: its data center positioning and future consumer-facing AI positioning are not yet seen by the market, but both are excellent future positions for Qualcomm. Qualcomm missed this wave somewhat—it's a latecomer—and its consumer-facing phones are affected by hardware price increases, so deliveries are poor. From the market setup, future potential, and risk/reward, it's quite good, and the valuation isn't expensive.

Mr. Z: But the market has been pretty harsh on hardware earnings lately. Qualcomm had a great report a week or two ago and still fell.

Investment Talk: The harshness over the past two months is really a positioning issue. The so-called "sell the news"—great earnings but the stock falls—means the market is over-positioned or in a distribution phase, with existing holders choosing to exit. But because of this, the entire semiconductor complex has been brought back to very low valuations. Investors should focus more on: which direction is this industry heading—improving, stagnant, or worsening? Are valuations at reasonable levels? If you have answers to these two questions, then one quarter's news or one earnings report is actually an opportunity.

7. Intel Is My Largest Position: Think of It as AMD Plus TSMC; Don't Use P/E on It

Mr. Z: Let's talk about Intel. It looks like the best case still needs to be repriced by the market: the 18A breakthrough, landing foundry customers, and Apple being very unhappy with TSMC this year. When can we see more explosive growth?

Investment Talk: I'm very bullish on Intel. After today's rise, it's become my largest position again, though I haven't added back the part I trimmed at $120. I bought it relatively early: at $20, its market cap was under $100 billion; now it's $500-600 billion. From an investment perspective, going from $500 billion to $1 trillion is a double. Is this name worth investing in? My answer is still yes. Can it double another 5x to $2.5 trillion? It's not entirely impossible, because you can think of its potential as AMD plus TSMC: it has its own design and also does foundry. The foundry is currently losing money, but as yields improve and customers come in, I believe by 2027 its foundry can turn positive.

The market, especially retail investors, has a big misconception about Intel's valuation: you can't just look at P/E. Its P/E of 60x or 70x looks expensive, mainly because the foundry is losing money. You absolutely cannot use P/E on a money-losing business—is its P/E infinite or zero? Foundry should be valued on P/S, applying a multiple to revenue; only the design and products segment should use P/E. Break it apart, and Intel is not expensive. Actually, I don't think many semiconductor names are expensive: Intel isn't expensive, Nvidia isn't expensive, Broadcom isn't expensive.

I'm not worried about Intel at all, because Lip-Bu Tan (Intel CEO) has proven his connections in the industry very well, and those connections have helped solve internal yield issues. Over the past two quarters of earnings, in the first quarter he said yields met company expectations but not his personal expectations; the next quarter he said yields had exceeded his own expectations. From PR self-promotion, landing customers, hiring, to yields—it's a complete 180-degree turn from other companies. Now it's also riding the CPU tailwind and the future AI agent tailwind. When will it explode? First from $500-600 billion to $1 trillion; after that, we'll see what multiple the market gives and how big the next outlook is. Fundamentals are continuously improving, and valuations will keep updating. With management like this and the entire industry's tailwind behind it, this company has very high certainty in my eyes, the valuation isn't expensive, and putting it as my largest position is very reasonable—it deserves it.

8. The Market Always Underestimates Nvidia: Depreciation Mismatch Is a False Proposition; Three-Year-Old Chips Pay Back in a Year

Mr. Z: Nvidia has been very cheap lately. What exactly is the market underestimating?

Investment Talk: The market can find many reasons to sell off, but we never know what the market is trading. You could say it's trading dissatisfaction with circular financing or other financing. Recently, it announced a partnership with six institutions to provide residual value guarantees for $500 billion of future GPUs. Ultimately, it comes down to implementation and demand: at current demand levels, A-series and H-series chips have no residual value risk at all. Chips from 3 or 5 years ago, at current rental prices, pay back in a year. This shows that 3 or 5 years ago, the entire market underinvested in AI infrastructure; current demand far exceeds the investment made back then, which is why demand far outstrips supply.

Anyone talking about whether Nvidia will face risks 3 or 5 years from now—nobody knows. You can only take it step by step. Over a year ago, people were talking about depreciation mismatch: GPUs should have a 3-year cycle, but Neoclouds (new cloud providers) and big clouds assumed 5-year depreciation, arguing that big clouds underestimated depreciation risk. Now the facts prove the opposite—big clouds were too conservative: at current on-demand rental demand, compute leased out by SpaceX pays back in a year. The concern from over a year ago about depreciation and payback period mismatch was a completely wrong conclusion. As for the B-series in hand and the upcoming Vera Rubin, nobody knows what rental prices will be in two or three years. But the previous deployment and current rental prices have already completely falsified that view.

The market's expectations for Nvidia have been wrong since 2023—wrong all along: the market never knows Nvidia's revenue growth for the next year. Both buy-side and sell-side just write down whatever guidance Nvidia gives; they never know the underlying demand. If they don't even know next year, who knows two or three years out? AI is the same—you can only have a vague evidence base and keep validating. Jensen Huang (NVIDIA CEO) gave 70% revenue growth guidance for the next fiscal year; the market expects 40%+. That just proves the market is underestimating again. At 17-18x forward earnings, even if the valuation doesn't re-rate, with 70% revenue growth and no operating leverage, EPS growth is also 70%. Even if the valuation contracts further, a 40% annual gain—EPS times P/E gives a 40% return—is still good. I see memory the same way: don't expect P/E to go very high. Assuming P/E stays flat, relying on EPS and fundamental growth for a 40% annual stock gain is already a pretty good result.

Many people ask if the market is running out of money. I never think the market is short of money; it's just in a process of continuous validation, which also means the market is healthy. June's mindless piling into chips and hardware and debating whether memory was cyclical—that was an unhealthy market. Now, with constant questioning, it's actually very healthy. As long as these companies keep executing, fundamentals keep monetizing, valuations are this low, and AI keeps landing and monetizing—what a great market.

9. The Three Big Clouds' Q3 Is an Open Book: When Fundamentals Are Accelerating, It's Absolutely Not a Selling Point

Mr. Z: On the CSP side, how do you view Meta, Microsoft, and Amazon? It feels like Q3 earnings should be good—last year's investments should all be realized.

Investment Talk: Google, Amazon, and Microsoft need to be viewed separately from Meta. The three big clouds' Q3 earnings are almost an open book: they will be very good and accelerating. Microsoft's Q2 earnings came out, and the stock only rose 4-5% after hours—the market reaction was quite slow. I said then, "Buy Microsoft"—I rarely say that so directly on X. The market completely underestimated Microsoft. The most underestimated point is its monetization in software and AI: Copilot's adoption grew 100% quarter-over-quarter to 30 million seats. At the time, the market doubted Copilot was doing terribly and nobody was using it—expectations were proven wrong. On the cloud side, Microsoft had explicitly said in previous earnings: it's not that they can't push Azure growth higher; it's that they internally consumed some compute. Cloud growth can be artificially managed—just allocate some compute to customers instead of internal use, and Azure growth naturally goes up. As a result, this quarter cloud growth went up, and Q3 guidance will accelerate further. Whether Microsoft, Google, or Amazon, cloud growth will continue to accelerate.

Microsoft's Q2 was 43%, next quarter's guidance is 45%, and it's very possible to hit 46% or 47%. When a company's fundamentals are accelerating, it's absolutely not the time to sell, because the market will extrapolate linearly. The market starts to worry when growth decelerates—say from 43% to 40% or 38%—still growing but at a smaller rate. Now it's going from 43% to 45%, and Q4 could be even higher, with margins still expanding. There's no reason to be bearish on these CSPs.

10. Anthropic Saved SpaceX and Meta's Compute

Investment Talk: More importantly, CSPs haven't yet benefited much from the rise in short-term rental prices. The ones who benefit immediately are Neoclouds: like SpaceX selling compute to Anthropic—they say it pays back in a year, even 9 months. CSPs sign relatively long-term contracts; long-term contracts are stable, but on-demand short-term rental prices only get reflected in long-term rentals when existing contracts expire and renew at on-demand pricing. Right now, on-demand prices are 2.5 to 3 times the 3-year contract prices of Neoclouds. So Neoclouds also benefit a lot: CoreWeave and Nebius both mentioned in earnings that newly signed contracts have higher margins than before, because capex is already behind them, and new contract prices are current. As long as on-demand keeps rising, new 1-year, 2-year, 3-year, 5-year contract prices will all rise, and payback periods get shorter. Capex is a rearview mirror—the money you invested is already behind you. You originally expected a 2.5-3 year payback period, not assuming prices would keep rising. But if on-demand keeps rising, it dramatically shortens the original payback period. That's why CoreWeave said new-year contract operating margins rose by 5 percentage points. For CSPs, margins are also a positive, just not reflected as much as for Neoclouds. Revenue growth accelerating and margins accelerating—there's nothing to worry about with these names.

Meta is different. It's not yet a cloud company. It may rent out compute in the future, but as a company monetizing AI, renting out compute isn't necessarily a good thing: renting out compute itself falsifies your original roadmap. If your own implementation is going great, why rent out compute? Of course, right now AI implementation is going well and there's a short-term offer, so they took it and sold some compute to Anthropic. Anthropic's contribution to the whole market is really huge: it raised prices, which is a positive for Neoclouds, CSPs, Meta, and SpaceX. SpaceX originally had excess compute and didn't know how to handle it—its internal monetization ability isn't strong. Same with Meta. Anthropic's monetization ability is very strong; it saved these two companies' excess compute by giving them an offer at a very good price. Meta also launched a personal agent today. Going forward, the test is still its AI implementation.

11. Amazon Only Makes $4 on a $100 API Call: Open-Source Models Hand Pricing Power Back to CSPs

Mr. Z: CSPs are now all adopting open-source models, distilling them and selling them. That also looks like one reason margins are improving.

Investment Talk: When I first invested in Amazon, I saw it as a service provider. From the start, it didn't intend to build its own model; it expected that whichever model wins will ultimately run on Bedrock. Early this year, I said Amazon's best monetization is Bedrock, and now Bedrock has become its monetization machine. The future trend is models becoming commodities. Right now, you could say closed-source models depend on CSPs, or CSPs depend on closed-source models. But from unit economics: for a $100 API call that ultimately lands, the bulk goes to the closed-source model—otherwise its gross margin couldn't be 80-90%. Out of $100, $90 goes to Anthropic, and the remaining $10 is what Anthropic pays CSPs as cost. Amazon's gross margin is 40%, so $6 is its original investment and cost—Amazon only makes $4. A $100 API call, reflected in Amazon's profit, is just $4. $90 of revenue is all captured by the closed-source model; pricing power is on the closed-source side.

With open-source models, it's different. Open-source models mostly come from China, and Chinese open-source models' monetization still has to rely on US CSPs, so pricing power shifts to the CSPs. And Amazon very early on let Chinese open-source models run on its platform. For CSPs, a flourishing ecosystem is the best outcome. As open-source model intelligence improves, it's a huge positive for CSPs and software. Earlier I said a $100 API call corresponds to an intelligence level of 130. That $100 is a cost eaten by software companies; they have to figure out how to pass that $100 to consumers, integrate it with products, and sell at higher prices. Now with open-source models, 3 to 6 months later, the same 130 intelligence, run on CSPs via open-source, might cost only $10 or $20. Software companies' costs drop from $100 to $20-30—costs are falling very fast. And they don't necessarily have to cut prices for customers, because customers don't know your costs; they only care about what you charge and whether you can deliver better monetization.

Gavin Baker in Silicon Valley recently said open-source models benefit the application layer and the infra layer. I completely agree—I held the same view before he said it: software companies' costs drop dramatically, they can experiment with different products, trial-and-error costs are low, development cycles are short, and if a product doesn't work, they kill it and move to the next. Taking it further: what used to cost $100 now costs $20, but software companies are still willing to spend $100—they'll just increase token volume by 5x. For CSPs, that's like doing 5 businesses instead of 1. Pricing power returns, volume goes up—definitely a positive for CSPs.

Mr. Z: Previously, the so-called bulk was captured by closed-source models. Now it looks like CSPs can win a round. How will the open-source vs. closed-source battle evolve? Right now, the more efficient approach is using closed-source models as the brain to direct, and open-source models for repetitive, mechanical tasks.

Investment Talk: That's already happening—it's not a prediction. Two or three months ago, Coinbase's CEO came out and said their spending is flat or growing very slowly, but token volume is rising exponentially, because they did model optimization behind the scenes. This is bound to happen: the hardest tasks go to closed-source, repetitive and relatively simple tasks go to open-source. Ultimately, token demand is higher than before, but spending isn't necessarily higher. From the CSP layer, Bedrock's position in the future is also to help enterprises optimize model usage—all models can be called on the platform, with recommendations based on needs, achieving the best optimization without a huge increase in total spending. The hardware side is also optimizing: High Bandwidth Memory (HBM) supply is constrained, and the new Vera Rubin has a lower HBM ratio. Qualcomm's data center roadmap is designed to solve future inference and agent inference demand and HBM constraints by separating decoding and prefill. Jensen Huang is also on this path—he acquired Groq (an AI chip company). Recently, Cerebras (an AI chip company)-Amazon and Amazon-Qualcomm partnerships are all along these lines. After spending too much money, everyone is feeling the pinch and will definitely choose a more rational path: model optimization on the consumer side, and hardware usage scenarios are also being separated—the division of labor among GPUs, XPUs (a type of processor), and Groq will become increasingly clear going forward.

12. Agent Implementation: 6 to 12 Months Out; Token Price Declines Are the Prerequisite for an Agent Explosion

Mr. Z: So where is the next AI implementation explosion? AI agents don't seem very general-purpose yet.

Investment Talk: There are already two examples in the last couple of days. Today Meta announced Muse, which does similar things to Gemini Spark—helping you order food, search, check your calendar, arrange things. The two have different strengths: Meta owns social media and social habits, so its implementation direction will be somewhat different. I also expect Apple's Siri to move toward consumer-facing agents. The other is OpenAI's Astra (GPT-6), which leans more toward business-facing agents. If the crayfish (a nickname for an AI agent product) from earlier this year was the first-generation agent, I expect that in the next 6 to 12 months, agent implementation will really start to materialize, because token prices have come down. Token price declines are the prerequisite for an agent explosion. In the first half of this year, agent-related names and narratives were basically not discussed, mainly because coding implementation was so good that all the market's compute was absorbed by coding. Now that open-source is catching up in intelligence and coding is gradually being commoditized, a lot of compute will be freed up, and prices will come down. Agent call volume rises exponentially compared to human calls. If prices don't come down, agents can't be implemented. Once intelligence goes up, problems that used to cost $1 to solve can now solve more and harder problems. So agent implementation must be accompanied by falling API prices and rising open-source model intelligence. Today Meta's Muse gives a 1 million token monthly limit, with subscription options to burn more. The fact that they dare to give you tokens to try suggests that token economics for agent implementation already makes sense. Plus, with a lot of compute coming online in 2027, I'm quite looking forward to real agent implementation in the next 6 to 12 months.

13. Nike and Lululemon's Common Problem: Too Complacent, Won't Admit Mistakes; Turnarounds Take at Least a Year

Mr. Z: You've also been looking at consumer names like Nike and Lululemon. Any comments?

Investment Talk: These two are traditional companies, not much related to AI. To get alpha in this market, you have to rely on AI, but I won't put everything into AI. Paying attention to and positioning in some traditional companies to diversify AI-related risk is necessary. Lululemon, Nike, and Starbucks from a year and a half ago—the three companies' biggest common problem: too complacent. Management completely failed to see changes in consumer demand; products had no innovation. Another point: they won't admit mistakes. Only when prices fell and it showed up in revenue did they realize they couldn't maintain their position by doing nothing. Once they realize it, it's a drag—product innovation and internal culture can't be fixed in a day or two. Turnarounds for these companies take at least a year. Starbucks has started its turnaround relatively successfully; recently, consumers have accepted its price increases.

Lululemon has tried to transform, but clearly not enough: management also mentioned that new products haven't been well received by the market. Its biggest problem is that it hasn't yet faced the impact of its biggest competitor, Alo (a yoga apparel brand), in the Chinese market. Alo only entered China in August, and China is Lululemon's best-growing market. In its best-growing market, a competitor that has already challenged its position in other markets has just arrived—you can only say the challenge in China is just beginning. New products haven't been accepted, which means at least the first wave of innovation failed. Fundamental change still needs time to verify. Nike is somewhat better: from other companies' earnings, it's been verified that its new products are well received. The problem is legacy old products—that's a huge problem. The shoe market, like overall US consumption, has consumers becoming very picky. High-end sales are actually not bad: high-income earners—the top 10%—account for 50% of US consumption. Good restaurants and Costco still have lines; mid-to-high-end consumption hasn't been affected at all. Brands like On Running (a running shoe brand) are more expensive than Nike, and consumers still buy. So Nike can't keep selling the same old Jordan brand; old products must be cleared at bigger discounts. New products have been accepted by the market. It's in the process of transformation—it's a mix. Looking forward, what matters isn't at what discount old products are sold, but whether new products can be accepted and their growth rate. Nike and Lululemon are fundamentally different.

14. Circle and Robinhood: Agent-to-Agent Is Still One Step Away

Mr. Z: After August 19, Bitcoin rallied, lifting crypto-native assets and crypto stocks. You said Circle and Robinhood have contributed noticeably to your portfolio. Let's talk about these two.

Investment Talk: I'm still underwater on Circle because I bought early and at a relatively high price. But this year, Circle has contributed quite a bit—it was big early in the year, then it fell from $130-140 to $60, and now it's back. Since August, it went from a 5% position to as high as 8%, all from price appreciation. The recent rebound in these two companies is definitely driven by Bitcoin, no doubt. But part of Circle's move may be overshadowed by Bitcoin's rise: earlier this year, when Circle rallied, the market was trading AI agents. It's just that after the crayfish came out, nobody used it—people found that spending $400-500 a month on API calls gave no return. The crayfish was the first product to carry the AI agent theme. But as I spent a lot of time discussing agent implementation, with API prices falling and intelligence rising, part of Circle's rise should also reflect the market's view on whether agent-to-agent payments, human-agent communication, and agent-merchant payments could use USDC. This has little to do with the company itself; it's the industry moving forward, and Circle had already positioned in this industry.

But Circle is positioned for agent-to-agent, and true agent-to-agent still needs time. We're just seeing the first step make sense: humans give agents instructions, and agents repeat what humans used to do. Today Meta's Muse announced a partnership with Stripe—you shop with Muse, and Muse pays via Stripe. That's still traditional payment; we haven't reached the stage of needing USDC payments. Right now it's just agent commerce—maybe step two or three, not yet agent-to-agent. It's one step away, but it's one step closer to agent-to-agent. The direction is correct, and it's a company building infrastructure. When it was in the $60s, I shared with many people to put a portion of their position in it, positioning for the future infrastructure layer. But management execution, compared to Robinhood, has a gap. Robinhood is basically the highest-execution among all crypto-related stocks. On-chain, many meme coins have done very well; their product execution is really strong, including prediction markets. So the market gives it a premium—valuation over 30x. Whether from a traditional brokerage or crypto perspective, it's not cheap, but from the CEO's execution, Robinhood deserves the premium. One has very strong management—no worries about missing hot trends; the other requires waiting—not yet at agent-to-agent. It's just a matter of buying some when it's cheap and trimming when sentiment is very high without clear implementation. Right now, Circle is nowhere near extremely high sentiment.

Mr. Z: On crypto itself, Bitcoin's outlook—how does it play out over the next three to six months?

Investment Talk: My personal view: this crypto wave was lifted by gold, and gold was lifted by real yields on Treasuries—that's the sequence. I think gold still has room to run, so Bitcoin also has room, and the whole market has room.

15. Tesla and SpaceX: Don't Bet Against Elon, But Don't Buy the Dream—Price Matters

Mr. Z: We forgot to talk about the most important ones—Tesla and SpaceX. What did you think of Tesla's Cybercab day?

Investment Talk: I thought the Robotaxi event was quite successful—successful in terms of publicity. It let many people know that the new Cybercab's biggest difference from Waymo is that it has no steering wheel, and the design is very striking—golden. For a company's product to be known, you have to create buzz and do marketing. Tesla itself doesn't do much marketing, so from the event itself, it was successful—it had gimmicks and generated buzz. But from a fundamental perspective, nothing changed: the autonomous driving that Cybercab can do, Model Y can also do. In Austin, they have both two-seaters and four-seaters. It didn't prove or disprove autonomous driving. I hold some Tesla, not a large position. I trimmed some a while ago and have been waiting for a good price opportunity to add back a portion, because I still believe widespread autonomous driving implementation will land on Tesla. It's just a matter of whether the price is right.

Mr. Z: What about our number one space stock, SpaceX?

Investment Talk: I think SpaceX is overvalued. From the results, the Cursor acquisition was well done—Cursor became a very good monetization channel for xAI. Originally, xAI clearly lacked a monetization channel. But most of Elon Musk's pitch comes from building data centers and selling compute. Selling compute itself isn't worth much; the market doesn't give great valuations to compute-selling companies unless you're a hyperscaler with value-added services. Otherwise, you can only reference Neocloud. Last quarter's earnings said the payback on selling compute is 9 months—less than a year. But a sub-one-year model isn't sustainable: if you build data centers, compute must be contracted long-term. Committing 8 GW of capacity yourself and signing short-term contracts is like shooting yourself in the foot. If future AI revenue relies entirely on selling compute, that's not a good business—it's asset-heavy, meaning future financing needs. Look at how painfully Oracle has had to finance. If Cursor takes off in coding, that could be a good growth point for SpaceX—it was a great acquisition. But I don't fully agree with the AI pitch. SpaceX is still expensive.

Mr. Z: Elon talks about a $30 trillion economy. Elon always loves to paint big pictures, but there's a saying in Silicon Valley: don't bet against Elon.

Investment Talk: "Don't bet against Elon" should be interpreted as don't bet against his vision: what you think is impossible, he can eventually achieve. But investing requires talking about price—you can't buy a dream. You have to distinguish between what's achievable and what's basically unachievable or very far off. Investing requires a time frame and a thesis—knowing under what conditions and after how long it can be falsified or confirmed, with a measurable standard. I agree with "Don't bet against Elon Musk," but that doesn't mean blindly investing. If SpaceX IPOs and you buy at $200+, you still have to talk about price.

Mr. Z: Give us an overview: from now until the Anthropic IPO is one phase, then the midterm elections, and after the elections—how will the next three to six months evolve? Reuters says the roadshow is delayed to late September, and the IPO should land in late October or early November. They're expanding the revolving credit facility (RCF) to $15 billion to bring in more US investment banks.

Investment Talk: For these companies going public—SpaceX and the two big model companies—you still need to be careful when investing, because they've all priced in the next two to three years or even further, and there's a huge overhang: primary market investors need to cash out. They have returns of dozens or even hundreds of times—they will definitely sell. This is a massive wealth transfer: when you buy these companies' stocks, money moves from secondary market participants to primary investors, who finally get their returns. From a liquidity perspective, these names are too large and will have an impact on the market—just like SpaceX's IPO drained Tesla. When the market looks for reasons to fall, these can be the reason; there will indeed be a sucking effect. But the main theme is still AI implementation. Including the midterm elections, I won't apply historical patterns—like the market doing great after midterms. I focus more on where short-term sentiment and participation are. If we enter the midterms with very high sentiment, then after the elections, it's very possible November will be down. Before August, many people thought September would fall; instead, the last couple of days software fell and hardware rebounded. Applying historical patterns isn't very meaningful. I follow sentiment and valuation: valuation is the higher-level judgment logic. If valuation is high and sentiment is high, trim a bit. Right now, valuation is low and sentiment is low—at the moment, I have nothing to worry about.

Mr. Z: At the end of the interview, anything you want to share or highlight?

Investment Talk: The last part of your question I find quite meaningful: if there's one sentence to share. Buffett said, "Buy when no one cares, sell when everyone's excited." It fits this year's market perfectly: early in the year, sentiment was very high; in March, the market fell and people worried about war—sentiment was very low. In June, everyone was euphoric about semis and hardware. In late July and late August, everyone was very pessimistic about semis. This year's market has tested us many times: when everyone's excited, will you sell? When no one cares, will you dare to buy? Can you actually implement this sentence in your own investing and trading? Right now, people might think it's cliché, but the experience Buffett left behind is actually quite useful.